FastAvatar: Fast 3D Avatar Reconstruction

• FastAvatar is a research domain focused on rapid, high-fidelity 3D avatar reconstruction that leverages feed-forward Gaussian Splatting and template-based residuals.
• It integrates pose-invariant identity coding from single images with transformer-based incremental fusion for unified multi-view and video inputs.
• Key trade-offs address balancing rendering speed and quality, enabling sub-10 ms inference or seconds-level refinement for animatable, photorealistic avatars.

Searching arXiv for papers on “FastAvatar” and closely related avatar reconstruction methods. FastAvatar denotes both a specific set of recent methods and a broader research objective: fast, high-fidelity 3D avatar reconstruction that remains animatable and view-consistent. In the recent literature, the name is attached to two feed-forward 3D Gaussian Splatting (3DGS) systems with distinct scopes. One FastAvatar reconstructs a pose-invariant face 3DGS from a single unconstrained face image by predicting residuals to a learned template in near-instant time; the other reconstructs an animatable canonical head 3DGS from a single image, sparse multi-view observations, or monocular video through a Large Gaussian Reconstruction Transformer and incremental Gaussian aggregation (Liang et al., 25 Aug 2025, Wu et al., 27 Aug 2025). Around these named systems sits a wider “fast avatar” lineage that includes motion-aware neural voxels, Instant-NGP-based dynamic neural fields, mesh-embedded Gaussians, few-shot feed-forward Gaussian avatars, and dynamic neural light fields (Xu et al., 2022, Jiang et al., 2022, Xiang et al., 2023, Wang et al., 2024, Nguyen et al., 14 May 2026).

1. Terminology and scope

In the supplied literature, “FastAvatar” is not a single historical architecture. It refers, first, to titled methods—specifically the 2025 papers “FastAvatar: Instant 3D Gaussian Splatting for Faces from Single Unconstrained Poses” and “FastAvatar: Towards Unified Fast High-Fidelity 3D Avatar Reconstruction with Large Gaussian Reconstruction Transformers”—and, second, to a broader design target repeatedly invoked by neighboring work: reducing fitting time while preserving animation control and photorealistic rendering (Liang et al., 25 Aug 2025, Wu et al., 27 Aug 2025, Xu et al., 2022).

Work	Input regime	Reported emphasis
FastAvatar (Liang et al., 25 Aug 2025)	Single face image from an arbitrary pose	Pose-invariant feed-forward 3DGS, single-pass inference at ≤10 ms, optional refinement in ~3 s
FastAvatar (Wu et al., 27 Aug 2025)	Single image, sparse multi-view observations, or monocular video	Unified feedforward LGRT, incremental reconstruction, seconds-level modeling

The ambiguity is substantive rather than merely lexical. The single-image FastAvatar is centered on a template-based residual formulation over a fixed canonical Gaussian set. The unified FastAvatar is centered on variable-length multi-frame aggregation, with a transformer that predicts per-frame canonical Gaussian sets and fuses them into a single animatable canonical representation. This suggests that the name marks a performance regime—fast reconstruction with high-fidelity rendering—as much as a single algorithmic family.

2. Single-image FastAvatar: template residuals and pose-invariant identity coding

The single-image FastAvatar formulates face reconstruction as residual prediction around a data-driven template 3DGS model rather than direct estimation of all Gaussian parameters. Template construction begins from multi-view optimized subjects in Nersemble, using FLAME initialization with two Gaussians per vertex; per-Gaussian parameters are then averaged across subjects to obtain a template with $K = 10{,}144$ Gaussians. Each Gaussian carries position $\mu_i \in \mathbb{R}^3$, covariance $\Sigma_i \in \mathbb{R}^{3 \times 3}$ via scale–rotation decomposition, opacity $\alpha_i \in [0,1]$, and spherical harmonics coefficients $c_i \in \mathbb{R}^{3 \times (L+1)^2}$ with degree $L=3$, yielding 48 appearance coefficients per Gaussian (Liang et al., 25 Aug 2025).

The core inference path is feed-forward. An encoder $g_\phi(I) = g_{\mathrm{MLP}}(g_{\mathrm{FR}}(I))$ maps an arbitrary-pose input image to a pose-invariant identity code $w$, where $g_{\mathrm{FR}}$ is an ArcFace-style face recognition backbone and $g_{\mathrm{MLP}}$ is a lightweight projection MLP. A shallow decoder $\mu_i \in \mathbb{R}^3$0, conditioned on the global identity code and a learned Gaussian embedding $\mu_i \in \mathbb{R}^3$1, predicts residuals $\mu_i \in \mathbb{R}^3$2, $\mu_i \in \mathbb{R}^3$3, $\mu_i \in \mathbb{R}^3$4, $\mu_i \in \mathbb{R}^3$5, and $\mu_i \in \mathbb{R}^3$6 for each Gaussian. These residuals are applied through constrained transforms: $\mu_i \in \mathbb{R}^3$7, $\mu_i \in \mathbb{R}^3$8, $\mu_i \in \mathbb{R}^3$9, $\Sigma_i \in \mathbb{R}^{3 \times 3}$0, and $\Sigma_i \in \mathbb{R}^{3 \times 3}$1. By predicting residuals rather than absolute parameters, the model keeps inference well-conditioned and fast.

Training is split into decoder pretraining and encoder alignment. First, learnable identity codes $\Sigma_i \in \mathbb{R}^{3 \times 3}$2 and Gaussian embeddings $\Sigma_i \in \mathbb{R}^{3 \times 3}$3 are optimized jointly with the decoder under rendering supervision:

$\Sigma_i \in \mathbb{R}^{3 \times 3}$4

Second, with decoder and subject codes frozen, the encoder is trained to regress arbitrary views of the same identity to the same latent code using

$\Sigma_i \in \mathbb{R}^{3 \times 3}$5

This division explicitly separates identity-code learning from view-invariant regression.

The latent space is structured enough to support direct operations. Identity interpolation uses $\Sigma_i \in \mathbb{R}^{3 \times 3}$6. Attribute editing uses a global SVM direction, $\Sigma_i \in \mathbb{R}^{3 \times 3}$7, where $\Sigma_i \in \mathbb{R}^{3 \times 3}$8 is the normalized linear SVM normal separating latent codes by attribute labels. The paper reports examples including neutral-to-disgust expression and short-to-long hair, and emphasizes that edits remain 3D-consistent across viewpoints.

Quantitatively, the model reports 21.26 dB PSNR, SSIM 0.90, LPIPS 0.20, and MAE 0.03 in single-pass mode, and 24.09 dB, SSIM 0.92, LPIPS 0.18, and MAE 0.02 with an optional refinement of about 300 iterations taking roughly 3 s. Runtime is reported as ≤10 ms for single-pass 3DGS prediction on an NVIDIA A100, with real-time rendering preserved after refinement. Ablations attribute stability and accuracy to the averaged template, cosine-aligned encoder training, and the default $\Sigma_i \in \mathbb{R}^{3 \times 3}$9 Gaussian budget; for example, $\alpha_i \in [0,1]$0 yields 17.65 dB at 9 ms, while $\alpha_i \in [0,1]$1 yields 22.26 dB at 29 ms.

3. Unified FastAvatar: Large Gaussian Reconstruction Transformer and incremental fusion

The second FastAvatar generalizes beyond single-image input and introduces a unified feedforward pipeline for a single image, sparse multi-view frames, or monocular video. Its central component is the Large Gaussian Reconstruction Transformer (LGRT), a VGGT-style transformer that registers and aggregates variable-length observations into a canonical 3DGS avatar. Rather than fixing the number of input views during training and inference, the model is explicitly trained to accept arbitrary subsets of 2–16 frames, enabling what the paper calls incremental reconstruction: more observations improve the canonical Gaussian model instead of being discarded (Wu et al., 27 Aug 2025).

Each frame $\alpha_i \in [0,1]$2 is encoded by DINOv2 to produce tokens $\alpha_i \in [0,1]$3. These are augmented with multi-granular guidance through

$\alpha_i \in [0,1]$4

where $\alpha_i \in [0,1]$5 denotes camera pose, $\alpha_i \in [0,1]$6 denotes FLAME expression coefficients, and $\alpha_i \in [0,1]$7 denotes head pose. Coarse 3D positional prompts derived from FLAME mesh vertices are injected inside frame-attention blocks. LGRT interleaves dual-stream DiT-style frame attention with global attention, producing registered tokens $\alpha_i \in [0,1]$8 that a shared two-layer GS Head maps to per-frame canonical Gaussian sets $\alpha_i \in [0,1]$9.

Fusion is representation-level rather than optimization-based:

$c_i \in \mathbb{R}^{3 \times (L+1)^2}$0

The fused canonical avatar is then driven by Linear Blend Skinning in FLAME space to produce expression-conditioned Gaussians and rendered by the 3DGS rasterizer $c_i \in \mathbb{R}^{3 \times (L+1)^2}$1:

$c_i \in \mathbb{R}^{3 \times (L+1)^2}$2

The paper states that there is no explicit merge/split operator; instead, the network learns canonical alignment and redundancy resolution through attention and supervision.

Two auxiliary training devices are central. The first is landmark tracking:

$c_i \in \mathbb{R}^{3 \times (L+1)^2}$3

using 68-point facial landmarks to regularize local geometry. The second is sliced fusion supervision: a single-frame prediction $c_i \in \mathbb{R}^{3 \times (L+1)^2}$4 and a fused multi-frame prediction $c_i \in \mathbb{R}^{3 \times (L+1)^2}$5 are both rendered to the same target, and both are penalized by pixel losses. The total objective is

$c_i \in \mathbb{R}^{3 \times (L+1)^2}$6

with $c_i \in \mathbb{R}^{3 \times (L+1)^2}$7, $c_i \in \mathbb{R}^{3 \times (L+1)^2}$8, $c_i \in \mathbb{R}^{3 \times (L+1)^2}$9, and $L=3$0.

Optimization uses AdamW for 124K iterations at constant learning rate $L=3$1, with bfloat16 and gradient checkpointing on 8×H100 GPUs for about 12 days. At inference, the same model accepts 1–16 frames. The reported quality-speed trade-off is explicit: with one frame, FastAvatar reports PSNR 25.24, SSIM 0.861, LPIPS 0.125, animation FPS about 389, and modeling time 1.28 s; with four frames, it reports PSNR 26.33, SSIM 0.875, LPIPS 0.111, animation FPS about 176, and modeling time 4.42 s. Larger-view configurations are reported at about 111 FPS with 8.77 s modeling time, and about 53 FPS with 17.70 s modeling time, with PSNR around 26.4 and SSIM around 0.87.

Ablations identify sliced fusion as the primary stabilizer for multi-frame registration. At 10K steps, removing sliced fusion yields L1 0.0433, PSNR 20.56, SSIM 0.851, and LPIPS 0.192, while the full system yields L1 0.0401, PSNR 21.25, SSIM 0.853, and LPIPS 0.183. Removing tracking yields L1 0.0403, PSNR 21.22, SSIM 0.852, and LPIPS 0.182. The reported interpretation is that sliced fusion mitigates ghosting and misalignment, while landmark tracking improves local facial geometry, especially around eyes and mouth.

4. Technical lineages in fast avatar research

The broader fast-avatar literature shows that speed has been pursued through several distinct representational strategies rather than one dominant formulation.

AvatarMAV replaces heavy MLP-based dynamic NeRF components with explicit voxel grids. It decouples canonical appearance from expression motion, representing the former with an appearance voxel grid plus a tiny two-layer MLP and the latter with motion-aware neural voxels generated by weighted concatenation of 32 motion voxel-grid bases, each semantically aligned one-to-one with the first 32 BFM expression bases. The same 3DMM expression coefficients $L=3$2 modulate both the parametric template and the motion voxel bases:

$L=3$3

Training converges in about 5 minutes in pure PyTorch, with near-convergence in about 2 minutes, and self-reenactment results report MSE 0.0014, PSNR 30.4, SSIM 0.96, and LPIPS 0.038 (Xu et al., 2022).

InstantAvatar accelerates monocular full-body avatar reconstruction through a different triad: Instant-NGP hash-grid encoding, Fast-SNARF inverse skinning, and dynamic empty-space skipping. The canonical field is queried after posed-space samples are inverse-mapped to canonical correspondences, and per-pose occupancy grids plus a normalized shared occupancy grid reduce ray-marching cost. The paper reports 130× faster convergence than prior monocular NeRF-based avatars, acceptable visual quality in about 10 seconds of training, strong results in about 60 seconds, and interactive animation at 15 FPS at 540×540 on a single RTX 3090 (Jiang et al., 2022).

FlashAvatar adopts mesh-guided 3D Gaussian Splatting. A uniform 3D Gaussian field is embedded on a FLAME mesh via UV sampling, and a dynamic offset MLP conditioned on the canonical surface position and expression code predicts residual position, rotation, and scale. Because large facial deformations are handled by the tracked mesh rather than by a learned deformation field, the Gaussian budget remains compact: about 13.5K Gaussians at default UV resolution. The paper reports 300 FPS at 512×512 on an RTX 3090, minutes-level reconstruction, and average metrics of MSE $L=3$4, L1 $L=3$5, PSNR 32.33, SSIM $L=3$6, and LPIPS $L=3$7 across tested videos (Xiang et al., 2023).

LightAvatar abandons volumetric and mesh rendering entirely in favor of a dynamic neural light field. It maps ray descriptors and a view-dependent expression embedding directly to RGB in a single forward pass, followed by a lightweight super-resolution module. The ray descriptor is formed by concatenating $L=3$8 sampled 3D points along the ray and applying positional encoding, while the expression embedding is computed as $L=3$9 from a local feature bank and spatial attention. Distillation from a pretrained NeRF-based teacher plus a warping field for real-data correction stabilizes training. The reported runtime is 174.1 FPS at 512×512 on an RTX 3090 with no customized optimization (Wang et al., 2024).

FFAvatar shifts the problem from per-subject optimization to generalizable few-shot reconstruction. Its Multi-View Query-Former fuses source images into a canonical Gaussian head, while a FLAME estimator predicts expression $g_\phi(I) = g_{\mathrm{MLP}}(g_{\mathrm{FR}}(I))$0, articulation $g_\phi(I) = g_{\mathrm{MLP}}(g_{\mathrm{FR}}(I))$1, and head pose $g_\phi(I) = g_{\mathrm{MLP}}(g_{\mathrm{FR}}(I))$2 directly from pixels, removing offline FLAME extraction. The reported deployment profile is about 2 seconds reconstruction without personalization, about 10 seconds with personalization, and about 49 FPS animation on a single NVIDIA A100 GPU, with a reported +5.5 PSNR gain over LAM on NeRSemble (Nguyen et al., 14 May 2026).

Across these systems, the shared objective is constant—reduce optimization latency and render at interactive rates—but the mechanisms differ: explicit voxelization, hash-grid neural fields, surface-anchored Gaussian splats, feed-forward template residuals, variable-length transformer aggregation, or single-pass neural light fields.

5. Performance regimes and speed–quality trade-offs

The two papers named FastAvatar report strong results, but they do so under different tasks, datasets, and input protocols. The single-image FastAvatar evaluates novel view synthesis on Nersemble by taking one of 16 views as input and evaluating on the remaining 15; the unified FastAvatar trains on NeRSemble with variable-length inputs and evaluates across single-view, sparse monocular, and multi-view settings, with additional evaluation on Ava256. This suggests that the reported metrics define task-specific operating points rather than a single universal ranking (Liang et al., 25 Aug 2025, Wu et al., 27 Aug 2025).

For the single-image system, the central trade-off is between pure feed-forward prediction and brief refinement. Single-pass inference yields 21.26 dB, SSIM 0.90, LPIPS 0.20, and MAE 0.03 at ≤10 ms. About 300 refinement iterations increase performance to 24.09 dB, SSIM 0.92, LPIPS 0.18, and MAE 0.02 in about 3 s. A second trade-off concerns Gaussian count: $g_\phi(I) = g_{\mathrm{MLP}}(g_{\mathrm{FR}}(I))$3 is the default balance at 10 ms, whereas $g_\phi(I) = g_{\mathrm{MLP}}(g_{\mathrm{FR}}(I))$4 improves quality modestly to 22.26 dB and SSIM 0.93 but increases runtime to 29 ms.

For the unified system, the central trade-off is observation count. With one frame, modeling time is 1.28 s and animation runs at about 389 FPS, with PSNR 25.24, SSIM 0.861, and LPIPS 0.125. With four frames, modeling time increases to 4.42 s and animation FPS falls to about 176, but PSNR rises to 26.33 and LPIPS drops to 0.111. Two larger-view configurations further reduce FPS to about 111 and about 53 while modestly improving fidelity. The paper explicitly presents this as a quality-speed-tunable paradigm.

The ablations in both papers locate the source of these gains in their architectural priors. In the single-image system, the averaged template gives the best stability and accuracy, and both SSIM loss in decoder training and cosine similarity in encoder training contribute measurable improvements. In the unified system, sliced fusion is the dominant factor preventing ghosting and misalignment, while tracking loss mainly refines local facial geometry.

A broader cross-paper pattern is also visible. Fast-avatar methods that rely on strong structural priors—FLAME correspondence, canonical templates, or mesh anchoring—tend to reduce per-subject optimization time most aggressively. Systems that directly regress a canonical avatar in one pass or a few seconds often sacrifice some flexibility under unusual hair, accessories, or lighting; systems that allow a short refinement stage or optional personalization recover additional high-frequency detail. This pattern is explicit in the single-image FastAvatar’s optional refinement and FFAvatar’s optional personalization.

6. Limitations, ambiguities, and future directions

A first source of confusion is terminological. Several papers explicitly note that they do not reference a specific prior named “FastAvatar,” yet position themselves as answering the same need: fast training or adaptation, high-fidelity rendering, and practical deployment (Xu et al., 2022, Nguyen-Phuoc et al., 2023, Jiang et al., 2022). The recent appearance of two separate methods titled FastAvatar therefore marks convergence on a problem formulation rather than resolution around one canonical solution.

A second misconception is that feed-forward avatar reconstruction eliminates dependence on proxy structure. In practice, the current FastAvatar systems still rely heavily on priors and preprocessing. The single-image method depends on a template built from multi-view optimized subjects with one-to-one Gaussian correspondences through FLAME initialization and VHAP-assisted tracking. The unified transformer method uses rough camera pose, FLAME expression, and head pose from multi-view FLAME tracking, plus 68-point landmarks. Neighboring methods often rely on 3DMM fitting, FLAME or SMPL parameters, or tracked meshes as well (Liang et al., 25 Aug 2025, Wu et al., 27 Aug 2025, Xiang et al., 2023).

Current limitations are correspondingly structured. The single-image FastAvatar reports reduced fidelity for long hair, fine strands, clothing, unusual lighting, and out-of-distribution accessories; it also notes demographic imbalance in training data as a possible source of bias. The unified FastAvatar identifies instability in the VGGT DPT-style tracking head, limited supervision from 68 landmarks, and reduced robustness under fast motion, self-occlusion, and harsh lighting variation. Related systems report similar issues under extreme expressions, inaccurate 3DMM fitting, far-from-frontal views, or sparse-view ambiguity (Liang et al., 25 Aug 2025, Wu et al., 27 Aug 2025, Xu et al., 2022, Nguyen et al., 14 May 2026).

The future directions proposed in the literature are consistent. The single-image FastAvatar points to explicit dynamic expression modeling, disentangled identity and expression codes, temporal consistency for video and talking avatars, and improved modeling of hair, clothing, and illumination. The unified FastAvatar points to variance-aware filtering and anisotropic blending for Gaussian fusion, improved handling of hair and accessories, and stronger occlusion robustness. FFAvatar similarly emphasizes richer facial actuation, dynamic hair and accessories, advanced lighting and shading models, and temporal coherence (Liang et al., 25 Aug 2025, Wu et al., 27 Aug 2025, Nguyen et al., 14 May 2026).

Taken together, FastAvatar is best understood as a rapidly consolidating research area organized around a shared engineering target: reconstruct a high-fidelity, animatable 3D avatar from weak or sparse observations in seconds or milliseconds rather than minutes or hours. The field has not converged on a single representation. Instead, it is converging on a small set of principles—strong structural priors, explicit canonicalization, lightweight rendering-friendly representations, and learned mechanisms for pose/expression disentanglement—that now define the fast-avatar problem space.